⚡ Image Downscaling Attacks, Open-source Benchmark for App Gen and more

⚡ Image Downscaling Attacks, Open-source Benchmark for App Gen and more

Issue #6 · Week of October 6

Joe Heitzeberg • Founder at AI Tinkerers • ⏱️ 1 min read

Creating space for leading builders to share ideas, grow, and make an impact.

We pulled the standouts from the last 2 weeks: agentic workflow orchestration and multimodal security are on people’s minds. Check TextEvolve (Nick Ryan, NYC) for automated LLM program discovery, Pipelex (Louis Choquel, Paris) for declarative PLX pipelines that turn natural requirements into tested production workflows, and Anamorpher (Kikimora Morozova, NYC) for image-downscaling prompt injections that expose preprocessing risks. These are real, vetted builds we’re still thinking about - read on.

CompileBench: Testing LLMs on Build Systems Piotr Grabowski Founding Engineer at Quesma 📍 AI Tinkerers Poland • Sep 18 Piotr Grabowski from Quesma presented CompileBench Eval: Do You Need AGI to Compile Google Chrome? He demonstrated an open-source benchmark that forces LLMs to build real projects from scratch in Docker via shell, from simple utilities to complex projects with many dependencies. The talk dug into results, model quirks, and the internals, including how long-running tasks expose toolchain quirks and log clutter. It matters to builders as a practical guide for model selection and tooling in production-like workflows. Read more → TECH STACK LLMsOpenRouterDockerPythonCursor PROJECT LINKS compilebench.com quesma.com

Anamorpher Downscale Attack Kikimora Morozova Associate Security Engineer at Trail of Bits 📍 AI Tinkerers - New York City • Sep 17 Kikimora Morozova from Trail of Bits presented Image Downscaling Attacks on Production AI Systems. She showed an adversarial image that reads normal at high res but reveals hidden prompts after bicubic downsampling, enabling data exfiltration via Gemini CLI and Zapier MCP. The setup hinged on Anamorpher, built with Suha Hussain, using a least-squares embedding workflow and pixel-level perturbation visuals. For builders, it’s a reminder to harden image preprocessing and auditing in multimodal systems. Read more → TECH STACK GPT-4LangChainStreamlitPythonOpenAI API PROJECT LINKS github.com trailofbits.com @kiki_morozova

TextEvolve: Auto LLM Tuning Nick Ryan ML Engineer at Nearist 📍 AI Tinkerers - New York City • Sep 17 Nick Ryan, a NYC-based ML engineer, presented How to remove yourself from the LLM design loop. He demonstrates TextEvolve, an LLM-driven tool that automates the iteration loop by generating optimized Python scripts, tests, and edge cases. The code is open source, and the workflow blends prompt-driven exploration with automated validation. Takeaway: this approach offers faster prototyping and more reliable deployments for builders who want reproducible, testable AI workflows. Read more → TECH STACK PythonClaudeGemini PROJECT LINKS github.com nickcdryan.com Video @nickcdryan

AlphaEarth Forest Loss Nicolas Schuldt Software Engineer at Neblina Labs 📍 AI Tinkerers - Paris • Sep 18 Nicolas Schuldt presented 'Predicting Forest Loss Using AlphaEarth Embeddings', a live demo pairing 64-dim AlphaEarth embeddings with WRI data via similarity-weighted kNN to forecast deforestation. The system analyzes 2017 patches, finds twins by cosine similarity, and outputs risk probabilities with confidence intervals. It relies on live ecosystem fingerprinting and a lightweight embedding workflow, with public GitHub for reproducibility. Nicolas, a hands-on builder from Ecuador (Hexay), showed that this approach is scalable for conservation analytics and that attendees (people loved it). Read more → TECH STACK Google Earth Enginedeck.glH3FastAPIPython PROJECT LINKS neblinalabs.com neblinalabs.com

AgentPay Framework James kanyiri CTO at Pocket Watch 📍 AI Tinkerers - Nairobi • Sep 25 James Kanyiri from Pocket Watch presented Payment integration for AI Agents, a framework letting AI agents call payments as composable tools. The project unifies transactional flows and is backed by PayLink’s API plus MCP AI integration, with support for M-Pesa, Airtel, KCB, and Equity. A runnable demo at paylink-platform.vercel.app and the public repo (paylinkmcp/paylink) let builders inspect and reuse. Audiences appreciated the agent-first abstraction, which lowers repeated integration and speeds deployment. Takeaway: a practical path to production-ready agent payments. Read more → TECH STACK PythonMCP PROJECT LINKS paylink-platform.vercel.app github.com

More Great Builds

Quick hits from the community — demos worth bookmarking:

SAM: Agentic Event Mesh

Rey Riel • AI Tinkerers - Montreal • Sep 24

Rey Riel from Solace presented Building Agentic Systems using Solace Agent Mesh. The talk walked through creating agents, gateway interfaces, installing plugins, and running the system with SAM’s event-driven ADK and native MCP connectors. SAM orchestrates multi-agent workloads in real time, enabling low-latency agent-to-agent communication and plug-in extensibility. It stands out for practical, production-ready design, and audiences liked its scalable, event-driven approach. It hints at broad business potential if deployed as a product for cross-team automation.

Loading tech tags...

github.comsolace.com

slimtwin: safe medical AI

Oskar Wójcikiewicz • AI Tinkerers Poland • Sep 18

Oskar Wójcikiewicz from slimtwin presented How Not to Kill Anyone: Safety Layers in Medical Reasoning, an early look at a platform that builds a digital health twin from data in smart scales, wearables, and bloodwork. The demo targets dietitians and shows real-time adaptation, with a data-extraction pipeline for messy inputs and safety layers delivering medical-grade insights to non-clinicians. Peers noted the safety-first approach and robust extraction. It signals practical, cautious health AI with real-world potential.

Loading tech tags...

tryslimtwin.com

Pipelex: Agent Builds Workflow

Louis Choquel • AI Tinkerers - Paris • Sep 18

Louis Choquel, Pipelex CTO, presented Agent Builds Production AI Workflow in 4 Minutes, showing an agent that turns natural language requirements into a production pipeline. The demo uses a declarative TOML DSL to define pipes and contracts, with a Python runtime orchestrating deterministic execution and dry runs. It features per-pipe model selection and an open-source cookbook for sharing workflows. Feedback hinted appreciation for a reproducible approach and the potential to cut maintenance and cost.

Loading tech tags...

pipelex.comgithub.comVideo

Transformer Lab: Local LLM Tuning

Deep Gandhi • AI Tinkerers - Calgary • Sep 23

Deep Gandhi presented Transformer Lab: Training LLMs and Diffusion Models Locally. The open-source desktop app lets anyone build, tune, and run LLMs on their own hardware without writing code. He demonstrated fine-tuning LLMs and evaluating them on a MacBook Air with no extra hardware, including checkpointing. Key tools include one-click fine-tuning, job tracking, and PyTorch/HuggingFace under the hood. Audiences noted its practicality, and the local, private ML angle fits the community’s push toward accessible, reproducible AI.

Loading tech tags...

github.comtransformerlab.ai

Hexagone AI: multimodal anonymization

Mariia Zameshina • AI Tinkerers - Paris • Sep 18

Mariia Zameshina from Hexagone AI presented Unlocking Sensitive Data for AI: Automated Multimodal Anonymization. The demo shows a platform that automates anonymization across text, images, and audio, with PDFs anonymized in seconds while preserving data utility and GDPR compliance. It offers a plug-and-play API and on-prem options for enterprise deployment. Audience feedback hinted at its practical value, and the approach points to safe, scalable AI data pipelines for RAG. It matters to builders as privacy-preserving tooling becomes essential.

Loading tech tags...

hexagone.aihexagone.ai

ORION: robot HRI demo

Miguel Angel Gonzalez Rodriguez • AI Tinkerers - Bogotá • Sep 25

Miguel Angel Gonzalez Rodriguez, a software developer at Pontificia Universidad Javeriana, presented ORION, a modular, low-cost differential robot for human-robot interaction supporting a social interaction robotic platform. ORION runs ROS 2 with micro-ROS on ESP32 for real-time motor control and publishes to ROS topics. The demo covered the embedded comms layer and how cloud-deployed LLM is used with a local vision model (to detect human emotion), together with local NLP models. Edge inference on ESP32 + Coral TPU was highlighted as a practical hack; code lives at https://github.com/Tesis-ORION.

Loading tech tags...

github.com

SAM: Event-driven Agentic System

Cyrus Mobini • AI Tinkerers - Montreal • Sep 24

Cyrus Mobini from Solace presented Building Agentic System using Solace Agent Mesh. SAM is an event-driven, multi-agent system that orchestrates agents via the ADK, with gateway interfaces and pluggable plugins. He demonstrated constructing agents, wiring a gateway, and running plugins to compose workflows. Survey vibes were positive, with listeners appreciating the hands-on approach. For builders, this points to scalable, low-latency agent orchestration that could inform future tooling.

Loading tech tags...

github.comsolace.com

MongoDB AI Composition

Ran Mizrahi • AI Tinkerers - New York City • Sep 17

Ran Mizrahi, founder and CEO of Bit Cloud, presented From Generation To Composition: Building Production-Grade Software with AI. He showed how composing prompts and attachments yields a production-ready Product Hunt clone, built with React, Node.js, and MongoDB, under a modular, Bit-driven architecture. The demo emphasizes maintainability, safety, and scalability in AI-enabled software, a message that resonated with survey feedback hinting strong interest in production-grade composition. It’s a practical glimpse of how composable AI pipelines can power real-world apps for builders.

Loading tech tags...

github.combit.cloud

SandboxJS: AI sandbox toolkit

James Murdza • AI Tinkerers - San Francisco • Sep 24

James Murdza presented SandboxJS: Terraform for AI Sandboxes, an open-source library for quickly building agents on any AI sandbox provider. SandboxJS provides a unified TypeScript API for Linux cloud sandbox providers and abstracts core sandbox functions like running code and building from a Dockerfile, supporting computer-use, coding, and reasoning agents. James, an integrations engineer and AI engineer from E2B, brings practical deployment insight. (People loved it) It’s a portable toolkit for reproducible AI sandbox workflows, with GitHub and npm distributions.

Loading tech tags...

github.comnpmjs.com

Browser Use: Auto-Guides

Faizan Khan • AI Tinkerers - Montreal • Sep 24

Faizan Khan of SlashML presented Automating Browser Usage with AI and provided a walkthrough of claude-code sdk, and how it works. The talk walked through AI-driven headless browsers like Browserbase and Browser Use, showing session replay and vision-based control that turn onboarding into living guides and use of lightweight agents that adapt to UI changes. This approach promises practical onboarding docs and a path to a product that auto-generates guides as apps evolve. Takeaway: it shows turning browser tasks into repeatable, teachable workflows for builders.

Loading tech tags...

github.comdocsalot.dev

Hexagone: Auto Anonymize

Romain Boudet • AI Tinkerers - Paris • Sep 18

Romain Boudet from Hexagone AI demonstrated Hexagone: Auto Anonymize for AI, a tool that automatically redacts personal data in text, images, and audio within PDFs while preserving data utility. It executes fast multimodal anonymization and outlines a reproducible workflow, with a GitHub-linked repo for others to reproduce and extend. As a hands-on founder with Python, data science, and iOS tooling, his demo highlighted privacy workflows builders can adapt today; a few audience nods hinted at the real-world utility.

Loading tech tags...

You are one of 95,000+ readers from Anthropic, OpenAI, Google, Microsoft, Meta, Apple, Amazon, Nvidia, Netflix, Stripe, Databricks, Snowflake, and others — spanning frontier labs, big tech, startups, and top universities.